databricks-genai-inference 0.2.2

Interact with the Databricks Foundation Model API from python

Databricks Generative AI Inference SDK (Beta)

The Databricks Generative AI Inference Python library provides a user-friendly python interface to use the Databricks Foundation Model API.

It includes a pre-defined set of API classes Embedding, Completion, ChatCompletion with convenient functions to make API request, and to parse contents from raw json response.

We also offer a high level ChatSession object for easy management of multi-round chat completions, which is especially useful for your next chatbot development.

You can find more usage details in our SDK onboarding doc.

[!IMPORTANT]
We're preparing to release version 1.0 of the Databricks GenerativeAI Inference Python library.

Installation

pip install databricks-genai-inference

Usage

Embedding

from databricks_genai_inference import Embedding

Text embedding

response = Embedding.create(
    model="bge-large-en", 
    input="3D ActionSLAM: wearable person tracking in multi-floor environments")
print(f'embeddings: {response.embeddings[0]}')

[!TIP]
You may want to reuse http connection to improve request latency for large-scale workload, code example:

with requests.Session() as client:
    for i, text in enumerate(texts):
        response = Embedding.create(
            client=client,
            model="bge-large-en",
            input=text
        )

Text embedding (async)

async with httpx.AsyncClient() as client:
    response = await Embedding.acreate(
        client=client,
        model="bge-large-en", 
        input="3D ActionSLAM: wearable person tracking in multi-floor environments")
    print(f'embeddings: {response.embeddings[0]}')

Text embedding with instruction

response = Embedding.create(
    model="bge-large-en", 
    instruction="Represent this sentence for searching relevant passages:", 
    input="3D ActionSLAM: wearable person tracking in multi-floor environments")
print(f'embeddings: {response.embeddings[0]}')

Text embedding (batching)

[!IMPORTANT]
Support max batch size of 150

response = Embedding.create(
    model="bge-large-en", 
    input=[
        "3D ActionSLAM: wearable person tracking in multi-floor environments",
        "3D ActionSLAM: wearable person tracking in multi-floor environments"])
print(f'response.embeddings[0]: {response.embeddings[0]}\n')
print(f'response.embeddings[1]: {response.embeddings[1]}')

Text embedding with instruction (batching)

[!IMPORTANT]
Support one instruction per batch Batch size

response = Embedding.create(
    model="bge-large-en", 
    instruction="Represent this sentence for searching relevant passages:",
    input=[
        "3D ActionSLAM: wearable person tracking in multi-floor environments",
        "3D ActionSLAM: wearable person tracking in multi-floor environments"])
print(f'response.embeddings[0]: {response.embeddings[0]}\n')
print(f'response.embeddings[1]: {response.embeddings[1]}')

Text completion

from databricks_genai_inference import Completion

Text completion

response = Completion.create(
    model="mpt-7b-instruct",
    prompt="Represent the Science title:")
print(f'response.text:{response.text:}')

Text completion (async)

async with httpx.AsyncClient() as client:
    response = await Completion.acreate(
        client=client,
        model="mpt-7b-instruct",
        prompt="Represent the Science title:")
    print(f'response.text:{response.text:}')

Text completion (streaming)

[!IMPORTANT]
Only support batch size = 1 in streaming mode

response = Completion.create(
    model="mpt-7b-instruct", 
    prompt="Count from 1 to 100:",
    stream=True)
print(f'response.text:')
for chunk in response:
    print(f'{chunk.text}', end="")

Text completion (streaming + async)

async with httpx.AsyncClient() as client:
    response = await Completion.acreate(
        client=client,
        model="mpt-7b-instruct", 
        prompt="Count from 1 to 10:",
        stream=True)
    print(f'response.text:')
    async for chunk in response:
        print(f'{chunk.text}', end="")

Text completion (batching)

[!IMPORTANT]
Support max batch size of 16

response = Completion.create(
    model="mpt-7b-instruct", 
    prompt=[
        "Represent the Science title:", 
        "Represent the Science title:"])
print(f'response.text[0]:{response.text[0]}')
print(f'response.text[1]:{response.text[1]}')

Chat completion

from databricks_genai_inference import ChatCompletion

[!IMPORTANT]
Batching is not supported for ChatCompletion

Chat completion

response = ChatCompletion.create(model="llama-2-70b-chat", messages=[{"role": "system", "content": "You are a helpful assistant."},{"role": "user", "content": "Knock knock."}])
print(f'response.text:{response.message:}')

Chat completion (async)

async with httpx.AsyncClient() as client:
    response = await ChatCompletion.acreate(
        client=client,
        model="llama-2-70b-chat",
        messages=[{"role": "system", "content": "You are a helpful assistant."},{"role": "user", "content": "Knock knock."}],
    )
    print(f'response.text:{response.message:}')

Chat completion (streaming)

response = ChatCompletion.create(model="llama-2-70b-chat", messages=[{"role": "system", "content": "You are a helpful assistant."},{"role": "user", "content": "Count from 1 to 30, add one emoji after each number"}], stream=True)
for chunk in response:
    print(f'{chunk.message}', end="")

Chat completion (streaming + async)

async with httpx.AsyncClient() as client:
    response = await ChatCompletion.acreate(
        client=client,
        model="llama-2-70b-chat",
        messages=[{"role": "system", "content": "You are a helpful assistant."},{"role": "user", "content": "Count from 1 to 30, add one emoji after each number"}],
        stream=True,
    )
    async for chunk in response:
        print(f'{chunk.message}', end="")

Chat session

from databricks_genai_inference import ChatSession

[!IMPORTANT]
Streaming mode is not supported for ChatSession

chat = ChatSession(model="llama-2-70b-chat")
chat.reply("Kock, kock!")
print(f'chat.last: {chat.last}')
chat.reply("Take a guess!")
print(f'chat.last: {chat.last}')

print(f'chat.history: {chat.history}')
print(f'chat.count: {chat.count}')